Underfitting
When a model is too simple or undertrained to even capture the pattern — bad everywhere, including training.
Think of it like
Trying to trace a curvy coastline with a single straight ruler: you miss no matter how you place it.
Example
A tiny model on a hard task might sit at 55% on both training and test data — it never had the capacity to learn it.
How it actually works
Underfitting is high error even on the training set: the model lacks the capacity, the features, or the training time to fit the data. The fix is the opposite of overfitting’s — more capacity, better features, longer training. The art is landing between the two.
For product teams
If a model is mediocre even on data it’s seen, the problem is capacity or training, not just tuning.
For engineers
High bias: the model class is too constrained to fit the underlying function.
Related
- Overfitting — The opposite is overfitting.
- Generalization — Both are failures of generalization.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome